--- name: voice-clone description: Builds a personal writing style guide (a "VOICE GUIDE") from 3–5 samples of the user's own writing, then writes or rewrites any text in that voice. Use when the user wants AI output to sound like them, says "write like me", "clone my voice", "learn my style", shares old emails/posts/newsletters as samples, or asks to apply their voice guide to a draft. --- # Voice clone Two modes: - **BUILD**: the user gives 3–5 writing samples. You produce a `VOICE GUIDE` they can save and reuse forever. - **WRITE**: the user gives a `VOICE GUIDE` plus a task or a draft. You write in that voice. If the user gives samples and a task in the same message, run BUILD first, show the guide, then run WRITE. --- ## BUILD mode ### Step 1: Check the samples Good samples are things the user wrote **without AI**, for real readers: emails to clients, Slack messages, old posts, a newsletter, a proposal, voice-note transcripts. Ideal total: 800–3,000 words across at least 2 formats. If you get fewer than 3 samples or under ~500 words total, say so and ask for more. If you suspect a sample was AI-written (heavy use of the patterns in the banned list (`banned-phrases.md`, shipped with the `de-slop` skill and at the root of this kit)), point it out and ask whether to exclude it. Building a voice from AI text clones the AI. ### Step 2: Measure (actually count, don't guess) For the combined samples, record: | Metric | How | |---|---| | Average sentence length | words ÷ sentences | | Rhythm range | shortest and longest sentence; % of sentences under 8 words | | Paragraph length | typical number of sentences per paragraph | | Contractions | frequent / occasional / never | | Person | I / we / you; how often they address the reader directly | | Punctuation habits | em dashes, parentheses, semicolons, exclamation marks, ellipses, colons: count each | | Capitalization quirks | lowercase starts, ALL CAPS for emphasis, etc. | | Openers | how their pieces usually begin (first 5 words of each sample) | | Closers | how they sign off or end | | Formatting | bullets, headings, bold, emoji: how much | ### Step 3: Extract the fingerprint From the samples, pull: 1. **Signature words and phrases** they repeat (with quotes from the samples). 2. **Words they never use** that AI loves (check the banned list against their text: what's absent is a rule). 3. **Humor and attitude**: dry, self-deprecating, blunt, warm, sarcastic? Quote one example. 4. **How they make a point**: story first? number first? opinion first? question? 5. **How they handle uncertainty**: hedge, admit it plainly, or commit anyway? 6. **Their specifics habit**: do they use numbers, names, places, tools? How often? 7. **Taboos**: anything they clearly avoid (jargon, emojis, exclamation marks, corporate words, swearing). ### Step 4: Output the VOICE GUIDE Use this exact template so the user can paste it into any chat or project instructions: ``` VOICE GUIDE: [NAME] Built from: [N] samples, [W] words, formats: [list]. Date: [date] 1. SOUND IN ONE LINE [e.g. "Talks like a tired but kind engineer explaining something to a friend over a beer."] 2. RHYTHM - Avg sentence: [N] words. Range: [min]–[max]. [X]% of sentences under 8 words. - Paragraphs: [N] sentences, often a one-line paragraph for emphasis: [yes/no] - Rule: [e.g. "After every long explanation, one short sentence that lands it."] 3. WORDS - Uses: [signature words/phrases, with quotes] - Never uses: [list] - Swears: [never / rarely / mild] - Contractions: [always / often / rarely] 4. PUNCTUATION & FORMAT - Em dashes: [count per 1,000 words → rule] - Exclamation marks: [rule] - Emoji: [rule] - Lists/bullets/bold: [rule] - Capitalization quirks: [rule] 5. HOW [NAME] MAKES A POINT - Opens with: [story / number / opinion / question] e.g. "[quote]" - Proves with: [numbers / examples / screenshots / stories] - Closes with: [CTA style / sign-off] e.g. "[quote]" 6. ATTITUDE - [e.g. "Admits mistakes first. Never sounds like a brochure. Opinions stated flat, no 'I think'."] - Example line: "[quote]" 7. DO / DON'T (the 10 rules a ghostwriter needs) DO: 1... 2... 3... 4... 5... DON'T: 1... 2... 3... 4... 5... 8. CALIBRATION PARAGRAPH [A 60–100 word paragraph you wrote in this voice on a neutral topic, for the user to approve or correct.] ``` ### Step 5: Calibrate Ask the user: "Does the calibration paragraph sound like you? Tell me one thing that's off." Update the guide with their correction. One round is usually enough. Remind them to save the final guide (in a Claude Project's instructions, a note, or a file named `voice-guide.md`). --- ## WRITE mode 1. Read the `VOICE GUIDE` fully. Follow it over any default style, including over the `de-slop` defaults where they conflict. 2. If the task needs facts you don't have (numbers, client names, results), leave `[ADD: ...]` gaps. Never invent them. 3. Draft. 4. Check against the guide line by line: - Is average sentence length within ±3 words of the guide? - Are the punctuation rules followed (count em dashes and exclamation marks)? - Any word from the "Never uses" list? Remove it. - Any phrase from the banned list (`banned-phrases.md`, shipped with the `de-slop` skill and at the root of this kit) that isn't in their signature words? Remove it. - Does it open and close the way they do? 5. Output the text, then one line: `Voice check: [what you matched] · [anything you couldn't match and why]`. --- ## Example (abridged) **Samples given:** 4 client emails and 2 LinkedIn posts by a freelance designer, ~1,400 words. **Part of the resulting guide:** ``` 1. SOUND IN ONE LINE Friendly, a bit dry, gets to the point by line two. Sounds like a designer who has been burned by vague briefs. 2. RHYTHM - Avg sentence: 11 words. Range: 2–31. 34% under 8 words. - One-line paragraphs for punchlines: yes. 3. WORDS - Uses: "honestly", "quick one", "here's what I'd do", "the boring answer is" - Never uses: leverage, solution, synergy, excited to, journey - Contractions: always 4. PUNCTUATION & FORMAT - Em dashes: 0. Uses parentheses for asides (≈4 per 1,000 words). - Exclamation marks: max 1 per email, only in the greeting. - Emoji: none in emails; one at the end of posts, sometimes. 7. DO / DON'T DO: start emails with "Quick one:" when it's a question. Give a recommendation plus a cheaper option. DON'T: open with a compliment. Use "I hope you're well". ``` **WRITE task:** "Tell the client the logo will be 3 days late." **Output:** > Quick one: the logo is going to land Thursday, not Monday. > > The honest reason: the second round of feedback changed the brief (wordmark instead of symbol), and I'd rather redo it properly than patch it. > > If Monday matters for the launch, I can send the wordmark in black only on Monday and the full set Thursday. Your call. > > Voice check: matched opener, parentheses, 2-option close, avg 12 words · no numbers to add. --- ## Limits - 3 samples gets you ~80% there. The last 20% comes from correcting the calibration paragraph and from the user editing outputs. Each correction should go back into the guide. - If samples are from different contexts (formal reports vs. casual DMs), build two guides or mark rules as "formal" / "casual". - It clones style, not knowledge. It can't know what the user thinks about a topic unless they say it.